MambaUNet3D: Combining Mamba State-Space Structures with 3D UNets to Segment Brain Tumors

The precise segmentation of brain tumors in 3D MRI presents a significant difficulty due to the very variable morphology and distribution of tumor subregions. Conventional CNN-based models like UNet excel at learning local spatial patterns but have difficulty capturing long-range contextual relationships, whereas Transformer-based methods typically necessitate extensive datasets and considerable computational resources. To mitigate these restrictions, we present MambaUNet3D, a hybrid architecture that combines 3D convolutional encoding with AxialMambaBlock3D modules to effectively model global dependencies without the burden of self-attention. The bottleneck additionally integrates an ASPP3D module to improve multiscale feature aggregation, and a skip-connected 3D decoder reinstates spatial resolution while maintaining anatomical fidelity. We assessed the model utilizing the BraTS 2020 dataset with a five-fold cross-validation methodology. Preprocessing encompassed N4 bias correction, modality-specific normalization, and balanced sampling to rectify class imbalance. The training curves exhibited consistent convergence, with training and validation loss, macro Dice, and accuracy closely aligned. Quantitative results demonstrated robust performance in both background and increasing tumor regions, achieving competitive Dice scores for edema and tumor core despite their intrinsic variability. Qualitative visualizations validated that MambaUNet3D precisely delineates tumor boundaries across Flair, T1ce, T1, and T2 modalities.